How Much Is a Token? (thefinancialengineer.substack.com)

Anthropic's Claude Devs announced a permanent 25% increase in Claude Code's standard weekly limit starting September 14, but this is notably less than the current promotion, which offers 50% more, potentially impacting user access to Claude Code. This shift highlights a broader issue in the AI and machine learning community regarding the concept of "tokens" as a currency for AI services. Since the introduction of a new tokenizer in Opus 4.7, the same input now yields roughly 30% more tokens, complicating the straightforward economics of token-based pricing and prompting a reevaluation of how AI services are billed. As companies increasingly rely on inference, which is anticipated to grow substantially—87 times by 2030 according to Dell—understanding costs beyond tokens becomes paramount. Inference can now differ greatly based on architectural choices and optimization strategies, meaning companies must carefully consider their inference providers to maintain competitive pricing structures. The recent acquisition of Hugging Face by NVIDIA for nearly $13 billion illustrates a pivotal industry shift, as NVIDIA gains control over both the compute resources and the platform where developers select models. This separation of layers further emphasizes the complexity of AI pricing systems, indicating that traditional token counting may not adequately reflect the true cost and valuation of AI services.